This article presents our study on task 10: Psycholinguistic conspiracy marker extraction and detection, which includes token-level extraction tasks and sentence-level conspiracy detection tasks. Focusing on conspiracy theory texts on social media, this paper proposes a classification method that combines semantic encoding with large language model reasoning and generation. Semantic features are extracted using DeBERTa-v3, and explanatory reasoning text is generated through ConspEmoLLM-v2. The two are then combined for classification, thereby enhancing the model’s ability to recognize implicit conspiratorial logic. For the extraction subtask, this paper provides systematic comparison results of several mainstream pre-trained models, mainly conducting baseline model comparisons and performance analysis.